EngineCL: Usability and Performance in Heterogeneous Computing

This research has been investigating an automatic and online crowd anomaly detection model by exploring a novel compound image descriptor generated from live video streams. A dual-channel convolutional neural network (DCCNN) has been set up for efficiently processing scene-related and motion-related crowd information inherited from raw frames and the compound descriptor instances. The novelty of the work stemmed from the creation of the spatio-temporal cuboids in online (or near real-time) manner through dynamically extracting local feature tracklets within the temporal space while handling the foreground region-of-interests (i.e., moving targets) through the exploration of Gaussian Mixture Model in the spatial space. Hence, the extracted foreground blocks can effectively eliminate irrelevant backgrounds and noises from the live streams for reducing the computational costs in the subsequent detecting phases. The devised compound feature descriptor, named as spatio-temporal feature descriptor (STFD), is capable of characterizing the crowd attributes through the measures such as collectiveness, stability, conflict and density in each online generated spatio-temporal cuboid. A STFD instance registers not only the dynamic variation of the targeted crowd over time based on local feature tracklets, but also the interaction information of neighborhoods within a crowd, e.g., the interaction force through the K-nearest neighbor (K-NN) analysis. The DCCNN developed in this research enables online identification of suspicious crowd behaviors based on analyzing the live-feed images and their STFD instances. The proposed model has been developed and evaluated against benchmarking techniques and databases. Experimental results have shown substantial improvements in terms of detection accuracy and efficiency for online crowd abnormal behavior identification.

/pdf/dual-channel-cnn-for-efficient-abnormal-behavior-5abj5jlzar.pdf

Dual-channel CNN for efficient abnormal behavior identification through crowd feature engineering

The parallel computation capabilities of modern GPU (Graphics Processing Unit) processors have attracted increasing attention from researchers and engineers who have been conducting high computational throughput studies. However, current single GPU based engineering solutions are often struggle to fulfill their real-time requirements. Thus, the multi-GPU-based approach has become a popular and cost-effective choice for tackling the demands. In those cases, the computational load balancing over multiple GPU “nodes” is often the key and bottleneck that affect the quality and performance of the runtime system. The existing load balancing approaches are mainly based on the assumption that all GPU nodes in the same computer framework are of equal computational performance, which are often not the case due to cluster design and other legacy issues. This paper presents a novel dynamic load balancing (DLB) model for rapid data division and allocation on heterogeneous GPU nodes based on an innovative fuzzy neural network (FNN). In this research, a 5-state parameter feedback mechanism defining the overall cluster and node performances is proposed. The corresponding FNN-based DLB model will be capable of monitoring and predicting individual node performance under different workload scenarios. A real-time adaptive scheduler has been devised to reorganize the data inputs to each node when necessary to maintain their runtime computational performances. The devised model has been implemented on two dimensional (2D) discrete wavelet transform (DWT) tasks for evaluation. Experiment results show that this DLB model has enabled a high computational throughput while ensuring real-time and precision requirements from complex computational tasks.

/pdf/a-fuzzy-neural-network-based-dynamic-data-allocation-model-1nrotg5anc.pdf

A Fuzzy Neural Network Based Dynamic Data Allocation Model on Heterogeneous Multi-GPUs for Large-scale Computations

Sigmoid: An auto-tuned load balancing algorithm for heterogeneous systems

Automatic video surveillance in public crowded places has been an active research area for security purposes. Traditional approaches try to solve the crowd behavior recognition task using a sequential two-stage pipeline as low-level feature extraction and classification. Lately, deep learning has shown promising results in comparison to traditional methods by extracting high-level representation and solving the problem in an end-to-end pipeline. In this paper, we investigate a deep architecture for crowd event recognition to detect seven behavior categories in PETS2009 event recognition dataset. More especially, we apply an integrated handcrafted and Conv-LSTM-AE method with optical flow images as input to extract a high-level representation of data and conduct classification. After achieving a latent representation of input optical flow image sequences in the bottleneck of autoencoder(AE), the architecture is split into two separate branches, one as AE decoder and the other as the classifier. The proposed architecture is jointly trained for representation and classification by defining two different losses. The experimental results in comparison to the state-of-the-art methods demonstrate that our algorithm can be promising for real-time event recognition and achieves a better performance in calculated metrics.

Real-time crowd behavior recognition in surveillance videos based on deep learning methods

Realizing the automated and online detection of crowd anomalies from surveillance CCTVs is a research-intensive and application-demanding task. This research proposes a novel technique for detecting crowd abnormalities through analyzing the spatial and temporal features of input video signals. This integrated solution defines an image descriptor (named spatio-temporal feature descriptor—STFD) that reflects the global motion pattern of crowds over time. A designed convolutional neural network (CNN) has then been adopted to classify dominant or large-scale crowd abnormal behaviors. The work reported has focused on: (1) detecting moving objects in online (or near real-time) manner through spatio-temporal segmentations of crowds identified by the similarity of group trajectory structures in the temporal space and the foreground blocks based on the Gaussian mixture model in the spatial space; (2) dividing multiple clustered groups based on the spectral clustering methods through treating image pixels from segmented regions as dynamic particles; (3) creating STFD descriptor instances by calculating corresponding attributes such as collectiveness, stability, conflict and crowd density for individuals (particles) in the corresponding groups; (4) inputting generated STFD descriptor instances into the devised CNN to detect suspicious crowd behaviors. For the test and evaluation of the devised models and techniques, the PETS database has been selected as the primary experimental data sets. Results against benchmarking models and systems have shown promising advancements of this novel approach in terms of accuracy and efficiency for crowd anomaly detection.

/pdf/towards-intelligent-crowd-behavior-understanding-through-the-3g3c6l9lrq.pdf

Towards Intelligent Crowd Behavior Understanding Through the STFD Descriptor Exploration

The powerful parallel computing capability of modern GPU (Graphics Processing Unit) processors has attracted increasing attentions of researchers and engineers who had conducted a large number of GPU-based acceleration research projects. However, current single GPU based solutions are still incapable of fulfilling the real-time computational requirements from the latest big data applications. Thus, the multi-GPU solution has become a trend for many real-time application attempts. In those cases, the computational load balancing over the multiple GPU nodes is often the key bottleneck that needs to be further studied to ensure the best possible performance. The existing load balancing approaches are mainly based on the assumption that all GPUs in the same system provide equal computational performance, and had fallen short to address the situations from heterogeneous multi-GPU systems. This paper presents a novel dynamic load balancing model for heterogeneous multi-GPU systems based on the fuzzy neural network (FNN) framework. The devised model has been implemented and demonstrated in a case study for improving the computational performance of a two dimensional (2D) discrete wavelet transform (DWT). Experiment results show that this dynamic load balancing model has enabled a high computational throughput that can satisfy the real-time and accuracy requirements from many big data processing applications.

/pdf/dynamic-load-balancing-on-multi-gpus-system-for-big-data-3hgy29822f.pdf

Dynamic load balancing on multi-GPUs system for big data processing

In order to solve the problem that the wrapped phase cannot be unwrapped effectively, this paper proposes an algorithm combining the branch-cut method and rhombus phase unwrapping algorithm for unwrap phase. The algorithm combines the advantages of the two methods, at first, the Goldstein's branch-cut method is used to find the residual points in the wrapping phase, then the rhombus phase unwrapping algorithm is used to unwrap the non-residual area phase, finally the phase of the residual area is obtained by using the cubic spline interpolation method, so as to ensure that the phase of each pixel can be unwrapped smoothly. The simulation results verify the feasibility and effectiveness of the algorithm, the combined phase unwrapping algorithm can effectively reduce the line and island phenomenon.

https://iopscience.iop.org/article/10.1088/1742-6596/1634/1/012068/pdf

An Algorithm Combining the Branch-Cut Method and Rhombus Phase Unwrapping Algorithm

Gaussian filtering has been extensively used in the field of surface metrology. However, the computing performance becomes a core bottleneck for Gaussian filtering algorithm based applications when facing large-scale and/or real-time data processing. Although researchers tried to accelerate Gaussian filtering algorithm by using GPU (Graphics Processing Unit), single GPU still fail to meet the large-scale and real-time requirements of surface texture micro- and nano-measurements. Therefore, to solve this bottleneck problem, this paper proposes a single node multi-GPUs based computing framework to accelerate the 2D Gaussian filtering algorithm. This paper presents that the devised framework seamlessly integrated the multi-level spatial domain decomposition method and the CUDA stream mechanism to overlap the two main time consuming steps, i.e., the data transfer and GPU kernel execution, such that it can increase concurrency and reduce the overall running time. This paper also tests and evaluates the proposed computing framework with other three conventional solutions by using large-scale measured data extracted from real mechanical surfaces, and the final results show that the proposed framework achieved higher efficiency. It also proved that this framework satisfies the real-time and big data requirements in micro- and nano-surface texture measurement.

Jun Lu

Papers

Towards Intelligent Crowd Behavior Understanding Through the STFD Descriptor Exploration

Dynamic load balancing on multi-GPUs system for big data processing

An Algorithm Combining the Branch-Cut Method and Rhombus Phase Unwrapping Algorithm

Multi-GPUs Gaussian filtering for real-time big data processing

An Improved Adaptive Window Stereo Matching Algorithm